Hugging Face's logo

Datasets:
SamuelChien821
/
dealbench-100

Tasks:
Question Answering
Languages:
English
ArXiv:
Tags:
agents
benchmarking
finance
tool-use
mcp
harbor
License:
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
channel: string
messages: list<item: struct<author: string, text: string>>
  child 0, item: struct<author: string, text: string>
      child 0, author: string
      child 1, text: string
terminal_growth: double
approved: bool
exit_multiple: double
entry_multiple: double
revision: string
wacc: double
to
{'approved': Value('bool'), 'entry_multiple': Value('float64'), 'exit_multiple': Value('float64'), 'revision': Value('string'), 'terminal_growth': Value('float64'), 'wacc': Value('float64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              channel: string
              messages: list<item: struct<author: string, text: string>>
                child 0, item: struct<author: string, text: string>
                    child 0, author: string
                    child 1, text: string
              terminal_growth: double
              approved: bool
              exit_multiple: double
              entry_multiple: double
              revision: string
              wacc: double
              to
              {'approved': Value('bool'), 'entry_multiple': Value('float64'), 'exit_multiple': Value('float64'), 'revision': Value('string'), 'terminal_growth': Value('float64'), 'wacc': Value('float64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

DealBench-100

DealBench-100 is a 100-task, deterministic investment-banking agent benchmark over ten synthetic transaction worlds. It tests source control, QoE normalization, trading comps, precedents, DCF, LBO, merger math, bid comparison, model-to-deck consistency, and launch approval.

Run 

harbor run -d blobfishai/dealbench-100-suite -a <agent> -m <provider/model>

Metric

The single metric is DealScore (0–100): discovery 15, model accuracy 25, decision 15, committed state 20, deliverable consistency 10, post-write readback 10, containment 5. Exact call order is not graded. Every point is executable; no LLM judge is called.

Release facts

  • 100 tasks; 10 synthetic project worlds; 10 workflow families
  • 26 agent-visible files per task across 10 native formats
  • 37 provider-shaped tools across 7 logical MCP servers
  • before/after state snapshots and full tool trajectories
  • 100/100 oracle strict passes, exact deterministic replays, and five negative-control families with zero false accepts
  • ranked rows are admitted only from complete, version-pinned, no-retry runs; inspect model-runs/ (and Harbor's model-runs.json) when present

Source

The complete synthetic world, task generator, deterministic verifier, release receipts, and model-run manifests are published at https://github.com/blobfishai/deal-agent-simulation/tree/main/benchmark/dealbench100.

All companies, bids, financials, approvals, messages and transactions are synthetic. This dataset is for agent evaluation and research; it is not financial or investment advice.

Public design anchors and clean-room boundary

DealBench-100 is independently authored. These public sources informed its release and evaluation shape:

The gated APEX-Agents dataset states that it is for evaluation only and forbids crawling/scraping and training use. It was not downloaded or scraped. No gated task, file, gold output, world snapshot, or trajectory was transformed or copied into DealBench. We used only the public benchmark descriptions and public illustrative sample to identify general desiderata: realistic professional outcomes, data-rich worlds, cross-application trajectories, before/after snapshots, and criterion-level grading.

The public Enterprise-Bench and ERP-Bench pages informed the use of noisy enterprise state, protocol-realistic interfaces, isolated task packages, and deterministic execution. Their tasks, worlds, assets, values, and solutions were not copied.

DealBench differs materially in content and evaluation: ten new synthetic companies and deal processes, new prompts and artifacts, provider-shaped closed-world tools, deterministic source/math/state verifiers, explicit collateral-damage checks, and no judge model.
Downloads last month
-
Total file size:
3.99 MB

Paper for SamuelChien821/dealbench-100